How to search on Oriane

oriane.xyz/how-to-search

Oriane indexes Instagram and TikTok at the frame level. Its AI analyzes what is on screen, what is spoken in audio, captions, hashtags, mentions, and context — all at once. This page explains how to use the search interface, what each filter does, how to design a search session that covers a topic completely, and what to do with the results.

0. The right mental model before you start

The deliverable is a paragraph for Oriane's AI prompt box

Oriane has a natural language prompt box at the top of the search page. The user describes the search they want; Oriane's internal LLM reads the description and configures the filters and tabs accordingly. The user no longer needs to manually pick filter types, set thresholds, or build tabs — that step is automated.

The catch: Oriane's filter-configuring LLM knows the filters but does not know the world. It does not know that MrBeast also runs Feastables, Beast Burger, Lunchly, and Beast Games. It does not know that "MrBeast" often comes out of the speech-to-text model as "Mr. Beast", "Mr Beast", or "Mister Beast". It does not know which hashtags a community uses, which crew members might be named instead of the main creator, or what the brand's packaging looks like.

That world knowledge is what you bring to the prompt. Research the topic, map every angle it appears under, and write a single natural language paragraph that contains all of it. Paste that paragraph into Oriane and you get a fully configured search session out of it.

The goal is full coverage, not a few good searches

The purpose of a search session on Oriane is to surface every relevant video that exists on the topic — not just the obvious ones. A search that is too narrow is a real loss: you permanently miss signal. A search that is too wide costs you nothing — false positives are sorted out after export by an AI reading every row.

Approach each topic with the explicit goal of covering all possible angles, all name variants, all associated people, all visual assets, all hashtags, all relevant caption vocabulary. You are not looking for the single best search query. You are building a system of complementary searches that together leave as little relevant content uncaptured as possible.

Always research the topic before building searches

Before touching the Oriane interface, do a web search on the topic you are tracking. This step is not optional — it is what separates a shallow search from one that covers the full spectrum of how people actually talk about the subject.

What to extract from that research:

Think in multiple searches, not one

A single Oriane search with many AND filters is precise but narrow. In most real use cases, you need to cover several distinct angles: verbal mentions, written mentions, visual appearances, related topic content, specific creator circles. Each angle often requires its own search.

Within a single search, all filters are AND conditions — every filter must match. To cover multiple angles without losing signal, you need multiple searches. Oriane handles this with a tab system: each tab is a self-contained search with its own filter combination, and tabs are unioned with OR logic. Up to 10 tabs in a session, and the whole session exports as one CSV. Use this fully. A well-designed search session for a non-trivial topic typically produces 5 to 10 tabs.

Discovery searches vs confirmation searches

There are two fundamentally different kinds of search sessions on Oriane, and confusing them is one of the most common mistakes.

Confirmation searches start from something you already know — a brand name, a product, a person, a hashtag, a known phrase — and look for content matching it. The vocabulary is the input. You feed Oriane the terms and it returns the videos.

Discovery searches start from a segment or a topic universe and aim to surface what is actually happening inside it — what patterns are emerging, what formulas are working, what creators are converging on. The vocabulary is the output. You ask Oriane for the universe of relevant videos, then analyze the export to discover the patterns.

Most brand monitoring, competitor tracking, IP protection, and creator vetting work is confirmation. But anything related to playbooks, hooks, viral mechanics, content formulas, format trends, and "what is working right now" is discovery. For these, you must not pre-filter on the patterns you expect to find — that biases the sample toward your priors and makes you blind to the formulas you have not named yet. Instead, scope the segment broadly, filter for performance, and let the analysis LLM derive the patterns from the raw transcripts and captions after export.

1. Filters

What filters are available

The Oriane search bar offers 12 filters that can be combined freely in a single search:

What each filter does

AI Vision

Describe what you want to see visually on screen. The input accepts natural language. Examples: "woman applying serum", "champagne being poured at a party", "blond girl brushing her hair".

A similarity threshold slider (default 70%) controls how strict the visual match is. The right setting depends on what you are searching for. For specific things like a product, a brand, or any search using text or an image reference, push the slider to around 90%. For globally recognized brands or products, this still returns a couple thousand results while cutting most of the noise. The default 70% will often blow past the 10,000-row export ceiling and bury genuine matches in false positives.

For generic visual descriptions like "blond woman", "guy with red shirt", or "someone brushing his teeth", leave the slider at 70%. Going lower is rarely useful since accuracy degrades meaningfully below this threshold. Only drop it further if results are sparse, and accept that the trade-off is more noise.

If the top results in the grid do not look like good matches and the sort is already set to "most similar visually", raising the slider will not help. You are already looking at the best matches the model found. Either revise the query, switch to an image reference, or fall back to Spoken Words and Caption filters.

You can also upload a reference image instead of typing. Oriane will find videos that contain the same or visually similar moment. Useful for finding copies, derivatives, or content built around a campaign visual. For topics with a strong visual identity — show posters, product shots, distinctive promotional images — uploading a reference image is often more effective than describing in text.

Important: AI Vision works well for universally recognized brands and highly distinctive visuals. Searching "Louis Vuitton bag" or "Mercedes car" or "Adidas shoes" will work because the model knows these brands visually. But niche brand names or less globally prominent people will not work well — the model has no reliable visual reference. For those, use Spoken Words or Caption instead. Also: because AI Vision relies on visual frame analysis, it can produce false positives that a text-based filter would have caught. When AI Vision is part of your search, lean on the transcript and caption data in your CSV to filter out visual mismatches during analysis.

Spoken Words

Filters by what is said aloud in the video audio — not what is written. This is one of the most powerful filters on the platform. Three modes:

Both Contains exactly and Doesn't contain exactly are multi-item fields. Add as many terms as you want, and toggle between any and all: "any" means a video matches if any one of the listed terms is spoken (OR across terms), "all" means every term must be spoken in the same video (AND across terms). For most brand or name searches, "any" is the right setting — list every spelling variant of the same name and let one of them match. Use "all" only when you genuinely need a video that mentions several distinct things at once.

Transcription accuracy warning. The speech-to-text model that generates transcripts is not perfect, especially for proper nouns — brand names, creator names, product names. A creator may clearly say "MrBeast" on camera, but the transcript may read "Mr. Beast", "Mr Beast", "Mister Beast", or a phonetic approximation of an unfamiliar brand name. Always run multiple spelling and formatting variants in separate Spoken Words searches. Do not assume a single exact-match search is sufficient. When in doubt, also pair every Spoken Words search with a Caption search — captions are typed by humans and are more reliably spelled.

Caption

Filters by the full written caption and description of the video. Uses the same three-mode logic as Spoken Words, including the multi-item any/all toggle on Contains exactly and Doesn't contain exactly. Caption is a broad and powerful signal: it includes everything written in the post, which means hashtags and mentions are also part of the caption text and will be caught by a Caption filter.

Caption and Spoken Words are distinct channels. A creator might write a brand name in their caption without saying it on camera, or say it on camera without writing it. Both angles must be covered. Always run dedicated Caption searches in addition to Spoken Words searches — do not assume that Spoken Words is sufficient.

Caption is also especially useful as a fallback. If you cannot find an account in the Mentions index, excluding it via Caption is a reliable alternative.

Language

Filter videos by the language spoken. Language splits are only useful when the search terms themselves differ by language — for instance, if a topic is described differently in French vs English, run a French-language search with French vocabulary. Do not split by language just to see results from a given country or region. If you are searching for a brand name or creator name that appears the same across languages (e.g., "Feastables", "MrBeast"), a single search without a language filter will return all languages at once, which is what you want. Let the analysis LLM sort, group, and quantify by language after export — it is faster, covers more ground, and avoids duplicating searches unnecessarily.

Platforms

Filter by social platform: TikTok, Instagram, or both. For maximum coverage, run both together. Only split by platform when you have a specific reason to isolate one.

Date

Filter by time period. The current data window is the last 3 months. It is not possible to search beyond this range. Also note that Oriane needs up to one week to index new videos, so very recent content may not yet appear — this is normal.

Mentions

Filter by accounts mentioned (@mentions) in the video post. Search an account name and Oriane suggests verified accounts with follower counts. Supports "any of these accounts are mentioned" and "any of these accounts are NOT mentioned". You can also paste a comma-separated list of accounts to add them in bulk.

Hashtags

Filter by hashtags used in the video post. Multiple hashtags in this filter are ORed by default — a video matches if it carries any one of them. Always research which hashtags are associated with a topic — there are often several variants (with and without accents, abbreviated forms, community-specific tags). Each distinct hashtag should be covered. If you are unsure which hashtags exist, search broadly first and observe what tags appear in the results.

Published by

Filter to videos published by specific accounts. You must search by the exact username or handle of the account — not the display name or brand name. A company named "Deel" may have the handle "@letsdeel", and searching "Deel" will not find it. If you do not know the handle, look it up on the platform directly, find it from a previous Oriane search result, or ask an LLM to find it with a web search.

Brands do not always use the same handle on Instagram and TikTok. Large corporations may also have multiple accounts split by language, region, or country. You can paste a comma-separated list of handles to search multiple accounts in one go — they are ORed together. Use NOT published by to filter out brand-owned content and isolate third-party UGC.

Note: Oriane does not index videos from accounts with fewer than 5,000 followers.

Followers

Filter creator accounts by follower count. Set a min and max, or use quick presets: 10K, 100K, 500K, 1M. Minimum indexed account size is 5,000 followers.

Views

Filter videos by view count range. Use to focus on high-performing content, or to find hidden gems with strong engagement despite modest reach.

Important: do not apply a views filter on specific brand or name searches. When you are searching for a precise term — a brand name, a product name, a creator's name — the result set is naturally bounded. You are unlikely to hit the 10,000-row export limit, and applying a minimum views threshold will silently discard relevant content for no benefit. Views filters should be reserved for broad, open-ended searches where the result volume is genuinely too large to export and analyze in full.

Engagement Rate

Filter by engagement percentage or total interactions (likes, comments, shares, saves). Toggle between percentage and interaction count depending on whether you want relative or absolute performance.

How searches combine — tabs, AND, OR

A search session is built from one or more tabs at the top of the search bar. Each tab is a self-contained search with its own combination of filters. The whole session exports as a single CSV.

Within a tab — AND between filters. Every filter you add narrows the result set. Spoken Words "hello" + Language "English" + Date "Last 30 days" returns videos that match all three at once. More filters means more precision; fewer filters means wider coverage.

Within a multi-item filter — any (OR) or all (AND). Several filters accept multiple values at once: Spoken Words and Caption (Contains exactly, Doesn't contain exactly), Hashtags, Mentions, Published by, Platforms. For Spoken Words and Caption, an explicit toggle switches between "any" (OR — match if any one term is present) and "all" (AND — every term must be present). For Hashtags, Mentions, Published by, and Platforms, multiple values are ORed by default. Most brand and name searches need OR across spelling variants, not AND — set "any".

Across tabs — OR between searches. Each tab runs as its own search; results are unioned and deduplicated. A session with one tab "Spoken Words: bonjour + Language: French" and a second tab "Spoken Words: hello + Language: English" returns every video matching either combination. The combined session exports as a single CSV — no need to export per tab and merge externally. Up to 10 tabs per session.

Describe a distinct angle when it requires a different combination of filters: a different filter type, a different filter mode, or a contradictory constraint. Examples: a Spoken Words search paired with a Caption search on the same name; a French-language search with French vocabulary alongside an English-language search with English vocabulary; an AI Vision visual search alongside a Hashtag search. Each angle becomes a tab in Oriane; together they form full coverage.

Do not describe separate angles for cases where the only difference is a value inside a single multi-item filter — list both values together in one filter description. And if the user wants to compare two distinct universes side by side rather than unite them, that should be run as two separate sessions with two separate CSVs, since the export does not track which tab a row came from.

2. How to design a search session

Step 1 — Research the topic

Before opening Oriane, do a web search. For any non-trivial topic, you should come out of this research with: every name or alias the topic is known by; key people associated with it; associated brands, platforms, and distribution context; visual assets (poster, logo, key imagery, distinctive scenes); related topics and adjacent conversations; known hashtags or creator communities.

Step 2 — Map every possible angle

List every distinct way the topic appears on social video. The goal here is exhaustiveness — write down every angle before deciding which ones to pursue:

Caption and Spoken Words always come in pairs. Never run one without the other for the same term. A creator may write a brand name in their caption without saying it on camera, or say it on camera without writing it. Two channels, two searches — always.

AI Vision for physical products. When a topic has a physical product with distinctive packaging, two visual angles matter: a specific packaging or product description (describe color, shape, characteristic markings, where it typically appears on screen — and if a clean product image is available, upload it directly in Oriane's AI Vision filter after the prompt runs); and a broad category-level visual angle (e.g., "burger on a tray", "chocolate bar wrapper") with a Views or Engagement Rate threshold to keep volume manageable.

Step 3 — Write the prompt

Combine the research and the angle map into a single natural language paragraph. This is the deliverable. The user pastes it into Oriane's AI prompt box; Oriane's filter-AI translates it into a configured search session.

You do not need to name filters explicitly ("use Spoken Words exactly"). Oriane's AI knows the filters. You just need to describe what to look for, where to look for it, and what to constrain. The job of the paragraph is to inject the world knowledge Oriane's AI is missing: name variants, associated entities, hashtags, accounts, visual descriptions, performance constraints.

What every paragraph must contain:

What to be explicit about for constraints:

Step 4 — Expectations, export, and analysis

Once the user runs the search, expect false positives. A name variant might catch unrelated content. An AI Vision direction might surface videos that look visually similar but are off-topic. A broad caption search might return posts where the keyword appears in a different context. This is normal and expected. The goal is not zero noise — it is zero missed signal.

The user exports the session as a single CSV; all tabs are unioned and deduplicated automatically. Up to 10,000 rows per export. The user then drops the file into Claude (or another model) with a prompt describing what counts as relevant and what to exclude. The model reads every caption and transcript and cleans the dataset in one pass. What feels like noise at the search stage disappears cleanly at the analysis stage.

3. Example prompts

Three worked examples covering the most common patterns. The text inside each block is exactly what gets pasted into Oriane's AI prompt box — nothing more.

Example 1 — Brand mention tracking (confirmation search)

User asked: "I want to find every video mentioning MrBeast or his associated stuff."

Research that fed the prompt: MrBeast also operates Feastables (chocolate), Beast Burger, Beast Games (Amazon Prime show), Lunchly (lunch kits), Beast Philanthropy, and ViewStats. His legal name is Jimmy Donaldson. His crew includes Chris Tyson, Karl Jacobs, Chandler Hallow, Nolan Hansen, and Tareq Salameh. Speech-to-text often renders "MrBeast" as "Mr. Beast", "Mr Beast", or "Mister Beast". Active hashtags: #mrbeast, #feastables, #beastgames, #beastburger, #lunchly. Official accounts: @mrbeast, @feastables, @beastburger, @beastgames, @lunchly.

Find every Instagram and TikTok video from the last 3 months that mentions MrBeast or anything associated with him. Look in both the audio transcript and the written caption for any of these spellings, since speech-to-text varies on proper nouns: MrBeast, Mr. Beast, Mr Beast, Mister Beast, and Jimmy Donaldson. Also include videos that mention any of his ventures by name, in either audio or caption: Feastables, Beast Burger, MrBeast Burger, Beast Games, Lunchly, Beast Philanthropy, ViewStats. Also include videos that name any of his crew in audio or caption: Chris Tyson, Karl Jacobs, Chandler Hallow, Nolan Hansen, Tareq Salameh. Include videos tagged with any of these hashtags: #mrbeast, #feastables, #beastgames, #beastburger, #lunchly. Include videos published by any of @mrbeast, @feastables, @beastburger, @beastgames, @lunchly, and videos that @-mention any of those accounts. Do not apply a language filter — these names appear the same across languages. Do not apply a views or engagement threshold — this is a name search, the result set is naturally bounded, and a threshold would silently drop relevant content.

Example 2 — Tracking a cultural release (confirmation search with visual angle)

User asked: "Track all social content around the new Dune Part Three release."

Research: title is also written Dune 3, Dune III, Dune: Part Three, and Dune Messiah. Directed by Denis Villeneuve. Main cast: Timothée Chalamet, Zendaya, Austin Butler, Florence Pugh, Rebecca Ferguson, Javier Bardem, Léa Seydoux. Distributed by Warner Bros and Legendary. Visually distinctive: sandworms in desert, stillsuits, the blue-eyed close-up, the Atreides emblem. Hashtags: #DunePartThree, #Dune3, #DuneMovie, #DuneMessiah, #Dune, plus actor-led tags like #TimotheeChalamet, #Zendaya. Accounts: @dunemovie, @warnerbros, @legendary.

Find every Instagram and TikTok video from the last 3 months related to the Dune Part Three film release. In both the audio transcript and the written caption, look for any of these title spellings: Dune Part Three, Dune 3, Dune III, Dune Messiah, Dune: Part Three. Also include videos that name the director — Denis Villeneuve, Villeneuve — or any of the main cast in audio or caption: Timothée Chalamet, Timothee Chalamet, Zendaya, Austin Butler, Florence Pugh, Rebecca Ferguson, Javier Bardem, Léa Seydoux. Include videos tagged with any of #DunePartThree, #Dune3, #DuneMovie, #DuneMessiah, #Dune, #TimotheeChalamet, #Zendaya. Include videos published by @dunemovie, @warnerbros, or @legendary, and videos that @-mention any of those accounts. Also include videos whose visuals show iconic Dune imagery — sandworms in desert sand, a character in a stillsuit, the Atreides emblem, or the distinctive glowing blue-eyed close-up — use a strict visual match threshold around 90% since this imagery is highly distinctive. Do not apply a language filter; the film name and cast names appear the same across languages. Do not apply a views threshold — this is a name-based search.

Example 3 — Hook pattern discovery (discovery search)

User asked: "What hook formulas are working right now for beauty content on TikTok?"

Research: this is a discovery search — the goal is to surface hook patterns that have not been named yet. The verbal channel must stay open. Scope the segment broadly through visuals and category vocabulary, then filter aggressively for performance and mid-tier creator size where formulas get tested.

Find high-performing beauty content on TikTok from the last 60 days for a hook pattern discovery study. Scope the segment broadly using visual descriptions of beauty content: woman applying makeup, woman doing a skincare routine, product swatch on hand, mascara application close-up, lipstick application, foundation routine, GRWM setup, before-and-after makeup transformation. Also include the segment via caption keywords: makeup, skincare, beauty, routine, GRWM, foundation, lipstick, mascara, concealer, blush. Include videos tagged with any of #beautytok, #makeuptok, #skincaretok, #grwm, #makeuptutorial, #beautyroutine. Filter to videos with at least 500K views OR engagement rate above 5%. Scope creator size to mid-tier accounts between 10K and 500K followers — this is the band where hook formulas get tested and iterated; mega-accounts tend to go viral on identity rather than format. Critical: do not apply any spoken words or caption filter on opening phrases, hook patterns, or known viral phrasings. Leave the verbal channel completely open. The point of this search is to discover hook formulas from the export, not to confirm hooks already known — pre-filtering would bias the sample and defeat the analysis.

4. Use case patterns

The patterns below describe which dimensions to cover in the paragraph for each common use case. The filter-level detail is what Oriane's AI will set up internally — you do not need to spell out filter names in the paragraph.

Finding brand mentions your tools missed

Have the paragraph ask for videos where the brand name is spoken aloud in audio but does not appear in the written caption. This isolates pure Shadow Reach — videos every text-based listening tool misses entirely. Include all spelling variants of the brand name on the spoken side, and instruct that the caption must not contain the same variants.

Researching a trend or content category

The paragraph should describe the visual context in plain language, plus broad vocabulary that captures the verbal dimension, plus relevant date scoping to see momentum. Research the topic first so the paragraph captures all the vocabulary people actually use.

Tracking a TV show, film, or cultural release

Research first: collect the full cast, the director, the distributor, the release platform, associated hashtags, and key visual assets. The paragraph should then cover every major name variant (show title, director, lead actors) across audio and caption with all spelling variants listed; a visual description of distinctive imagery with a strict match threshold called out where the imagery is iconic; every known hashtag; a related-topic angle if the show connects to a broader cultural conversation; and an explicit instruction not to apply a language filter.

Finding creators for a partnership

Think beyond brand names. Have the paragraph describe the content context visually in plain language, plus a follower range and an engagement rate minimum. The results are creators who already produce relevant content organically.

Vetting a creator before signing

Have the paragraph ask for every video published by the creator's account — by exact handle. Do not add any other filters; retrieve the full indexed history as-is. Give Claude the export with your specific red flag criteria for analysis.

Competitor intelligence

The paragraph should cover three distinct angles: videos published by the competitor's official accounts (to audit their own content strategy), videos that @-mention the competitor's accounts (to find third-party creators talking about them), and videos that mention the competitor's brand name in audio transcript or caption (with spelling variants). All three angles in one paragraph — Oriane's AI will split them into separate tabs.

Campaign measurement

Write one paragraph for the pre-campaign window and one for the post-campaign window, each scoped with the corresponding date range. Use the same core search content in both. Within each paragraph, cover both branded activation (specific campaign hashtags) and organic halo (brand name in audio or caption without the campaign hashtag) so the export lets the analysis LLM measure both effects.

Content playbooks, viral hooks, and format research

This is a discovery use case. The mistake to avoid: do not include hook phrases you already know ("nobody talks about", "POV you just", "wait for it") anywhere in the paragraph. That biases the sample toward your priors. The hooks worth surfacing are the ones nobody has named yet.

The correct approach: define the segment, not the formula. Scope the segment broadly through every channel except hooks (visual context, category vocabulary, community hashtags, optionally a handful of top accounts). Filter for performance — minimum view count (typically 500K or 1M+ for absolute virality), or engagement rate above 5% for hidden hits punching above their weight. Optionally scope to mid-tier creators (10K to 500K followers) where formulas get tested. Explicitly forbid pre-filtering on hook phrases.

After export, hand the dataset to Claude with a discovery prompt — make sure the export includes the Audio Transcript (Pro) column. Frame the analysis as pattern derivation, not pattern matching: extract the first 10 to 15 words of each transcript, cluster by structural pattern without a pre-existing taxonomy, and return the formula in plain language with performance stats for each cluster.

5. Exporting data

After any search session, click the export button to download a CSV of the results. The export covers all tabs in the session unioned and deduplicated by video — you do not export per tab. Select which columns to include before downloading. Up to 10,000 rows per export.

Available export columns:

The file is UTF-8 encoded. When parsing with Python or pandas, use encoding="utf-8-sig" to avoid BOM issues on the first column. Engagement Rate values are decimals — multiply by 100 to get the percentage.

6. Analyzing with AI

Once you have a CSV, drag it into Claude, ChatGPT, Gemini, or Copilot alongside a prompt. The AI reads every row — captions, transcripts, engagement data, creator details — and generates a structured intelligence report.

Oriane's prompt library covers the most common report types:

Each prompt includes full instructions for noise filtering, data cleaning, analysis framework, and HTML report design. Ready to copy and paste directly.

Prompt library: oriane.xyz/prompt-library

Browse example reports built from real Oriane data — Shadow Reach analyses, brand intelligence reports, and trend reports across categories — to see what the output looks like before running your own search.

Example reports: oriane.xyz/explore

7. Step-by-step guides by use case

For a walkthrough of the full process for each specific use case — from search setup to insight — see:

Oriane — AI Video Intelligence · oriane.xyz